This is a merge I made of sldr's photorealistic NSFW lora. Please check out and support his work.<a target="_blank" rel="ugc" href="https://civarchive.com/models/704013/photorealistic-nsfw">https://civarchive.com/models/704013/photorealistic-nsfw</a><h3 id="intent-fdj0snoj0">Intent</h3>My goal in making this merge was to create a flexible nudity lora that can be used with various other style and character loras. If baked into Flux Dev, it could also be used as a starting point to train other NSFW concepts without having to start over from scratch each time.I found that V1 was broad and flexible while V2 had greater quality but limited scope. I aimed to capture the flexibility and quality of both versions within a single lora. I believe I accomplished that fairly well.<h3 id="method-4z8a25mx9">Method</h3>This is a 50/50 merge between V1 and V2 of sldr's lora. After merging, I downsized the rank from 64 to 16. I believe this is what gives the merge its flexibility. When used with character loras of a similar rank (8-16 works best), there seems to be little to no conflict between loras. I have found that in general, you want your lora rank to be as low as possible without noticeable loss of quality. The lower the lora rank, the less likely it will be to conflict with other loras, and the more loras you can use at once.All example images were generated at 1024x1024 or 832x1216 using Flux Dev fp16 and then upscaled by 1.5 at 0.55 denoise. Generation parameters were dpmpp2, beta, 25 steps, and sigmas multiplied by 0.98. The multiplication of sigmas is roughly equivalent to adding extra noise and seems to help with out of focus backgrounds. Recommended lora strength is 1.15 (the details sections says 1.2, that is rounded).I also use separate prompts for clip L and T5. For clip L, I use several keywords. For T5, natural language.<h3 id="conclusion-7cwlwz24n">Conclusion</h3>The merge seems to have accomplished its stated goals. When used in conjunction with a character lora, it not only matches the face of a character, but also the body type. It even roughly approximates their bust size. It also seems to be able to handle a wide variety of settings, not being constrained to indoor/outdoor/day/night etc...The beauty of this lora is that it can be used with other NSFW loras to get a desired effect. For example, if you want larger breast sizes, throw in Large Natural Breasts by Felldude at 0.5 strength and reduce this lora's strength by the same amount. You should get a pretty seamless merge of the two.This is pretty close to a perfect bypass of Flux's censorship while being as lightweight as possible so as not to interfere with other loras. I believe it should be baked into Flux Dev and used as a starting point for all other NSFW training.

Photorealistic NSFW Merge

SldrMix_rank16.safetensors

v1.0

This is a merge I made of sldr's photorealistic NSFW lora. Please check out and support his work.<a target="_blank" rel="ugc" href="https://civitai.com/models/704013/photorealistic-nsfw">https://civitai.com/models/704013/photorealistic-nsfw</a><h3 id="intent-fdj0snoj0">Intent</h3>My goal in making this merge was to create a flexible nudity lora that can be used with various other style and character loras. If baked into Flux Dev, it could also be used as a starting point to train other NSFW concepts without having to start over from scratch each time.I found that V1 was broad and flexible while V2 had greater quality but limited scope. I aimed to capture the flexibility and quality of both versions within a single lora. I believe I accomplished that fairly well.<h3 id="method-4z8a25mx9">Method</h3>This is a 50/50 merge between V1 and V2 of sldr's lora. After merging, I downsized the rank from 64 to 16. I believe this is what gives the merge its flexibility. When used with character loras of a similar rank (8-16 works best), there seems to be little to no conflict between loras. I have found that in general, you want your lora rank to be as low as possible without noticeable loss of quality. The lower the lora rank, the less likely it will be to conflict with other loras, and the more loras you can use at once.All example images were generated at 1024x1024 or 832x1216 using Flux Dev fp16 and then upscaled by 1.5 at 0.55 denoise. Generation parameters were dpmpp2, beta, 25 steps, and sigmas multiplied by 0.98. The multiplication of sigmas is roughly equivalent to adding extra noise and seems to help with out of focus backgrounds. Recommended lora strength is 1.15 (the details sections says 1.2, that is rounded).I also use separate prompts for clip L and T5. For clip L, I use several keywords. For T5, natural language.<h3 id="conclusion-7cwlwz24n">Conclusion</h3>The merge seems to have accomplished its stated goals. When used in conjunction with a character lora, it not only matches the face of a character, but also the body type. It even roughly approximates their bust size. It also seems to be able to handle a wide variety of settings, not being constrained to indoor/outdoor/day/night etc...The beauty of this lora is that it can be used with other NSFW loras to get a desired effect. For example, if you want larger breast sizes, throw in Large Natural Breasts by Felldude at 0.5 strength and reduce this lora's strength by the same amount. You should get a pretty seamless merge of the two.This is pretty close to a perfect bypass of Flux's censorship while being as lightweight as possible so as not to interfere with other loras. I believe it should be baked into Flux Dev and used as a starting point for all other NSFW training.

Files with SHA256: 0525f9f3d08ce542cc262b4d30ffe7509d784290322230e415c0e76a7679c80b